gumble-max-vllm-experiment / plot_ranks_histogram.py
royrin's picture
Upload folder using huggingface_hub
d16e9ef verified
"""
Plot histogram of ranks computed with >= comparison or from logit ranks.
"""
import pickle
import numpy as np
import matplotlib.pyplot as plt
from pathlib import Path
import argparse
from tqdm import tqdm
from datetime import datetime
def compute_ranks_gte(data, sigma):
"""Compute ranks using >= comparison from Gumbel scores."""
ranks = np.zeros(len(data), dtype=np.int32)
for i, item in enumerate(tqdm(data, desc=f"Computing ranks (>=, sigma={sigma})")):
sampled_score = item['sampled_gumbel_scores'][sigma]
top_k_scores = item['top_k_gumbel_scores'][sigma]
# Rank = number of tokens with score >= sampled_score
num_higher_or_equal = np.sum(top_k_scores >= sampled_score)
ranks[i] = num_higher_or_equal
return ranks
def extract_logit_ranks(data):
"""Extract logit ranks directly from data."""
ranks = np.zeros(len(data), dtype=np.int32)
for i, item in enumerate(tqdm(data, desc="Extracting logit ranks")):
ranks[i] = item['logit_rank']
return ranks
def extract_gumbel_scores(data, sigma):
"""Extract Gumbel scores for sampled tokens."""
scores = np.zeros(len(data), dtype=np.float32)
for i, item in enumerate(tqdm(data, desc=f"Extracting GLS scores (sigma={sigma})")):
scores[i] = item['sampled_gumbel_scores'][sigma]
return scores
def plot_scores_and_ranks_comparison(gls_scores, gls_ranks, logit_ranks, output_path, sigma):
"""Create 3-panel comparison: GLS scores, GLS ranks, and logit ranks."""
fig, axes = plt.subplots(1, 3, figsize=(20, 6))
total = len(gls_scores)
# Left panel: GLS scores histogram
ax = axes[0]
bins = np.linspace(-20, 0, 101) # 100 bins from -20 to 0
counts, bin_edges = np.histogram(gls_scores, bins=bins)
percentages = (counts / total) * 100
# Plot as bars
bin_centers = (bin_edges[:-1] + bin_edges[1:]) / 2
bin_width = bin_edges[1] - bin_edges[0]
# Filter out zero counts for log scale
non_zero_mask = percentages > 0
ax.bar(bin_centers[non_zero_mask], percentages[non_zero_mask],
width=bin_width, edgecolor='black', alpha=0.7)
ax.set_xlabel('GLS Score', fontsize=12)
ax.set_ylabel('Percentage (%) - Log Scale', fontsize=12)
ax.set_title('GLS Scores Distribution')
ax.set_yscale('log')
ax.grid(alpha=0.3, which='both')
ax.set_xlim(-20, 0)
# Middle panel: GLS ranks histogram (grouped: 0, 1, 2, ..., 19, 20+)
ax = axes[1]
# Group ranks: 0-19 individually, 20+ together
grouped_ranks = []
grouped_labels = []
for i in range(20):
count = np.sum(gls_ranks == i)
grouped_ranks.append(count)
grouped_labels.append(str(i))
# Add 20+ group
count_20_plus = np.sum(gls_ranks >= 20)
grouped_ranks.append(count_20_plus)
grouped_labels.append('20+')
grouped_percentages = (np.array(grouped_ranks) / total) * 100
# Filter out zero counts for log scale
non_zero_mask = grouped_percentages > 0
x_positions = np.arange(21)[non_zero_mask]
ax.bar(x_positions, grouped_percentages[non_zero_mask], edgecolor='black', alpha=0.7, width=0.8)
ax.set_xlabel('GLS Rank (0=highest)', fontsize=12)
ax.set_ylabel('Percentage (%) - Log Scale', fontsize=12)
ax.set_title('GLS Ranks Distribution')
ax.set_yscale('log')
ax.set_xticks(range(21))
ax.set_xticklabels(grouped_labels, rotation=0, fontsize=10)
ax.grid(alpha=0.3, which='both', axis='y')
# Right panel: Logit ranks histogram
ax = axes[2]
unique_ranks, counts = np.unique(logit_ranks, return_counts=True)
percentages = (counts / total) * 100
ax.bar(unique_ranks, percentages, edgecolor='black', alpha=0.7, width=0.8)
ax.set_xlabel('Logit Rank (0=highest)', fontsize=12)
ax.set_ylabel('Percentage (%) - Log Scale', fontsize=12)
ax.set_title('Logit Ranks Distribution')
ax.set_yscale('log')
ax.grid(alpha=0.3, which='both')
plt.tight_layout()
plt.savefig(output_path, dpi=150, bbox_inches='tight')
print(f"\nSaved comparison plot to {output_path}")
plt.close()
# Print statistics
print("\n" + "="*80)
print(f"GLS Scores, GLS Ranks, and Logit Ranks Comparison (sigma={sigma})")
print("="*80)
print(f"Total items: {total:,}")
print(f"\nGLS Scores statistics:")
print(f" Min: {gls_scores.min():.4f}")
print(f" Max: {gls_scores.max():.4f}")
print(f" Mean: {gls_scores.mean():.4f}")
print(f" Median: {np.median(gls_scores):.4f}")
print(f" Std: {gls_scores.std():.4f}")
print(f"\nGLS Ranks statistics:")
print(f" Min: {gls_ranks.min()}")
print(f" Max: {gls_ranks.max()}")
print(f" Mean: {gls_ranks.mean():.2f}")
print(f" Median: {np.median(gls_ranks):.2f}")
print(f"\nLogit Ranks statistics:")
print(f" Min: {logit_ranks.min()}")
print(f" Max: {logit_ranks.max()}")
print(f" Mean: {logit_ranks.mean():.2f}")
print(f" Median: {np.median(logit_ranks):.2f}")
print(f"\nTop 10 most common GLS ranks:")
unique_gls_ranks, gls_counts = np.unique(gls_ranks, return_counts=True)
top_10_idx = np.argsort(gls_counts)[-10:][::-1]
for idx in top_10_idx:
rank = unique_gls_ranks[idx]
count = gls_counts[idx]
pct = 100 * count / total
print(f" Rank {rank:3d}: {count:6,} ({pct:6.2f}%)")
print(f"\nTop 10 most common logit ranks:")
unique_logit_ranks, logit_counts = np.unique(logit_ranks, return_counts=True)
top_10_idx = np.argsort(logit_counts)[-10:][::-1]
for idx in top_10_idx:
rank = unique_logit_ranks[idx]
count = logit_counts[idx]
pct = 100 * count / total
print(f" Rank {rank:3d}: {count:6,} ({pct:6.2f}%)")
def plot_rank_histogram(ranks, output_path, title_suffix, max_rank_plot=30):
"""Create histogram of ranks."""
fig, axes = plt.subplots(2, 2, figsize=(14, 10))
fig.suptitle(f'Rank Distribution ({title_suffix})', fontsize=16, fontweight='bold')
# Get rank distribution
unique_ranks, counts = np.unique(ranks, return_counts=True)
total = len(ranks)
# Separate data for plotting (ranks 1-max_rank_plot) and ranks > max_rank_plot
plot_mask = unique_ranks <= max_rank_plot
plot_ranks = unique_ranks[plot_mask]
plot_counts = counts[plot_mask]
# Count ranks > max_rank_plot
higher_ranks_mask = unique_ranks > max_rank_plot
count_above_max = np.sum(counts[higher_ranks_mask])
# Plot 1: Bar plot of rank distribution
ax = axes[0, 0]
ax.bar(plot_ranks, plot_counts, edgecolor='black', alpha=0.7, width=0.8)
ax.set_xlabel('Rank', fontsize=12)
ax.set_ylabel('Count', fontsize=12)
ax.set_title(f'Rank Distribution (Linear Scale, ranks 1-{max_rank_plot})')
ax.grid(alpha=0.3)
# Add text annotation for ranks > max_rank_plot
if count_above_max > 0:
pct_above = 100 * count_above_max / total
ax.text(0.98, 0.98, f'Rank > {max_rank_plot}: {count_above_max:,} ({pct_above:.2f}%)',
transform=ax.transAxes, ha='right', va='top',
bbox=dict(boxstyle='round', facecolor='wheat', alpha=0.5), fontsize=10)
# Add percentages as text labels for top ranks
for rank, count in zip(plot_ranks[:5], plot_counts[:5]):
pct = 100 * count / total
ax.text(rank, count, f'{pct:.2f}%', ha='center', va='bottom', fontsize=9)
# Plot 2: Bar plot with log scale
ax = axes[0, 1]
ax.bar(plot_ranks, plot_counts, edgecolor='black', alpha=0.7, width=0.8)
ax.set_xlabel('Rank', fontsize=12)
ax.set_ylabel('Count (log scale)', fontsize=12)
ax.set_yscale('log')
ax.set_title(f'Rank Distribution (Log Scale, ranks 1-{max_rank_plot})')
ax.grid(alpha=0.3)
# Plot 3: Percentage distribution
ax = axes[1, 0]
plot_percentages = (plot_counts / total) * 100
ax.bar(plot_ranks, plot_percentages, edgecolor='black', alpha=0.7, width=0.8)
ax.set_xlabel('Rank', fontsize=12)
ax.set_ylabel('Percentage (%)', fontsize=12)
ax.set_title(f'Rank Distribution (Percentages, ranks 1-{max_rank_plot})')
ax.grid(alpha=0.3)
# Plot 4: Cumulative distribution (use all ranks for this)
ax = axes[1, 1]
percentages = (counts / total) * 100
cumulative_pct = np.cumsum(percentages)
# Only plot up to max_rank_plot
plot_cumulative = cumulative_pct[plot_mask]
ax.plot(plot_ranks, plot_cumulative, linewidth=2, marker='o', markersize=6)
ax.set_xlabel('Rank', fontsize=12)
ax.set_ylabel('Cumulative Percentage (%)', fontsize=12)
ax.set_title(f'Cumulative Rank Distribution (ranks 1-{max_rank_plot})')
ax.axhline(y=95, color='red', linestyle='--', alpha=0.5, label='95%')
ax.axhline(y=99, color='orange', linestyle='--', alpha=0.5, label='99%')
ax.legend()
ax.grid(alpha=0.3)
plt.tight_layout()
plt.savefig(output_path, dpi=150, bbox_inches='tight')
print(f"\nSaved histogram to {output_path}")
plt.close()
# Print statistics
print("\n" + "="*80)
print(f"Rank Statistics ({title_suffix})")
print("="*80)
print(f"Total items: {total:,}")
print(f"\nRank distribution (ranks 1-{max_rank_plot}):")
for rank, count in zip(plot_ranks, plot_counts):
pct = 100 * count / total
print(f" Rank {rank:3d}: {count:6,} ({pct:6.2f}%)")
# Print ranks > max_rank_plot
if count_above_max > 0:
print(f"\nRanks > {max_rank_plot}:")
pct_above = 100 * count_above_max / total
print(f" Total count: {count_above_max:6,} ({pct_above:6.2f}%)")
print(f" Individual ranks:")
for rank, count in zip(unique_ranks[higher_ranks_mask], counts[higher_ranks_mask]):
pct = 100 * count / total
print(f" Rank {rank:3d}: {count:6,} ({pct:6.2f}%)")
print(f"\nCumulative distribution:")
cumsum = 0
for rank, count in zip(unique_ranks, counts):
cumsum += count
pct = 100 * cumsum / total
if rank <= max_rank_plot or rank == unique_ranks[-1]: # Print up to max_rank and the last rank
print(f" Rank ≤ {rank:3d}: {cumsum:6,} ({pct:6.2f}%)")
def main():
parser = argparse.ArgumentParser(
description="Plot histogram of ranks computed with >= comparison or from logit ranks"
)
parser.add_argument(
"--input",
type=str,
required=True,
help="Path to all_prompts.pkl file"
)
parser.add_argument(
"--plot-type",
type=str,
choices=["detailed", "comparison"],
default="comparison",
help="Type of plot: 'detailed' (4-panel rank histogram) or 'comparison' (2-panel GLS vs ranks) (default: comparison)"
)
parser.add_argument(
"--rank-type",
type=str,
choices=["gumbel", "logit"],
default="logit",
help="Type of rank to plot for detailed view: 'gumbel' or 'logit' (default: logit, only used with --plot-type detailed)"
)
parser.add_argument(
"--sigma",
type=float,
default=1.0,
help="Sigma value to use for GLS scores (default: 1.0)"
)
parser.add_argument(
"--output",
type=str,
default=None,
help="Output path for histogram (default: auto-generated based on plot type)"
)
parser.add_argument(
"--max-rank-plot",
type=int,
default=30,
help="Maximum rank to plot in detailed view (default: 30)"
)
args = parser.parse_args()
# Load data
print(f"Loading data from {args.input}...")
with open(args.input, 'rb') as f:
data = pickle.load(f)
print(f"Loaded {len(data):,} items")
input_path = Path(args.input)
output_dir = input_path.parent
if args.plot_type == "comparison":
# Extract GLS scores, GLS ranks, and logit ranks
gls_scores = extract_gumbel_scores(data, args.sigma)
gls_ranks = compute_ranks_gte(data, args.sigma)
logit_ranks = extract_logit_ranks(data)
# Generate output filename
if args.output is None:
datestr = datetime.now().strftime("%Y%m%d_%H%M%S")
output_path = output_dir / f"scores_vs_ranks_comparison_sigma{args.sigma}_{datestr}.pdf"
else:
output_path = Path(args.output)
# Plot comparison
plot_scores_and_ranks_comparison(gls_scores, gls_ranks, logit_ranks, output_path, args.sigma)
else: # detailed
# Compute ranks based on type
if args.rank_type == "gumbel":
ranks = compute_ranks_gte(data, args.sigma)
title_suffix = f"Gumbel scores, sigma={args.sigma}"
default_filename = f"ranks_histogram_gumbel_sigma{args.sigma}.pdf"
else: # logit
ranks = extract_logit_ranks(data)
title_suffix = "Raw logit ranks (0=highest logit)"
default_filename = "ranks_histogram_logit.pdf"
# Generate output filename
if args.output is None:
output_path = output_dir / default_filename
else:
output_path = Path(args.output)
# Plot histogram
plot_rank_histogram(ranks, output_path, title_suffix, args.max_rank_plot)
if __name__ == "__main__":
main()